Error estimates with explicit constants for Sinc quadrature and …conf.nsc.ru/files/conferences/scan2012/139602/Okayama-scan2012.pdf · 1 Sinc quadrature and Sinc inde nite integration

Error estimates with explicit constantsfor Sinc quadrature and Sinc indefinite

integration over infinite intervals

Tomoaki OKAYAMA (Hitotsubashi Univ.)

September 27, 2012

SCAN’2012

1 Sinc quadrature and Sinc indefinite

integration

1 Sinc quadrature and Sinc indefinite integration

Base: Sinc approximation on R

F (x) ≈N∑

j=−N

F (jh)Sinc(x/h− j)

Sinc(x) =sin(πx)

πx

-0.2 0

0.2 0.4 0.6 0.8

1 1.2

-10 -5 0 5 10

x

Sinc(x)=sin(πx)/(πx)

By integrating both sides, we can derive approximations for

• definite integrals (Sinc quadrature)∫∞−∞ F (x) dx ≈

∑Nj=−N F (jh)

h︷︸︸︷∫∞−∞ Sinc(x/h− j) dx

• indefinite integrals (Sinc indefinite integration)∫ τ

−∞ F (x) dx ≈∑N

j=−N F (jh)∫ τ

−∞ Sinc(x/h− j) dx

SCAN’2012 2 / 28


Important facts on two-step approximation

Sinc quadrature∫∞−∞ F (x) dx ≈ h

∑Nj=−N F (jh)∫ ∞

−∞F (x) dx ≈ h

∞∑j=−∞

F (jh) discretization

(accurate if the integral is over R)

≈ h

N∑j=−N

F (jh) truncation

(accurate if |F (x)| decays quickly as x→ ±∞)

Question When either the two “if” is NG ...?

SCAN’2012 3 / 28


Four typical cases that Stenger [2] consideredLet f be a smooth and bounded function

1. I1 = (−∞, ∞)

∫ ∞

−∞

f(t)

1 + t2dt polynomial decay

2. I2 = (0, ∞)

∫ ∞

0

f(t)

1 + t2dt polynomial decay

3. I3 = (0, ∞)

∫ ∞

0

e−tf(t) dt exponential decay

4. I4 = (a, b)

∫ b

a

f(t) dt (not decay)

SCAN’2012 4 / 28


Variable transformations for the 4 casesSingle-Exponential transformation x = ψSEi(x) (Stenger [2])

I1 = (−∞,∞) F (t) =1

1 + t2f(t)

I2 = (0,∞) F (t) =1

1 + t2f(t)

I3 = (0,∞) F (t) = e−tf(t)

I4 = (a, b) F (t) = f(t)

SCAN’2012 5 / 28


Variable transformations for the 4 casesSingle-Exponential transformation x = ψSEi(x) (Stenger [2])

I1 = (−∞,∞) F (t) =1

1 + t2f(t)

t = ψSE1(x) = sinh x

I2 = (0,∞) F (t) =1

1 + t2f(t)

t = ψSE2(x) = ex

I3 = (0,∞) F (t) = e−tf(t)

t = ψSE3(x) = arcsinh( ex)

I4 = (a, b) F (t) = f(t)

t = ψSE4(x) =b−a2

tanh(x2) + b+a

2

SCAN’2012 5 / 28


SE-Sinc quadrature (case 2)

∫I2

F (t) dt =

∫ ∞

−∞F (ψSE2(x))ψ

′SE2(x) dx SE transformation

≈ h∞∑

j=−∞F (ψSE2(jh))ψ

′SE2(jh) discretization

(accurate because the integral is over R)

≈ h

N∑j=−N

F (ψSE2(jh))ψ′SE2(jh) truncation

(accurate because the integrand decays O( e−|x|))

SCAN’2012 6 / 28


To make the truncation error further smallerDouble-Exponential transformation x = ψDEi(x) (Takahasi–Mori [3])

I1 = (−∞,∞) F (t) =1

1 + t2f(t)

t = ψDE1(x) = sinh[(π/2) sinhx]

I2 = (0,∞) F (t) =1

1 + t2f(t)

t = ψDE2(x) = e(π/2) sinhx

I3 = (0,∞) F (t) = e−tf(t)

t = ψDE3(x) = ex−exp(−x)

I4 = (a, b) F (t) = f(t)

t = ψDE4(x) =b−a2

tanh(π sinhx2

) + b+a2

Decay rate: O( e− exp(|x|))

SCAN’2012 7 / 28


Error analysis for SE/DE-Sinc quadrature

Case 1

∫ ∞

−∞

√1 + tanh2(arcsinh(t)/2)

1 + t2dt = 4arcsinh(1)

There exists a constant C independent of N such that

|Error (SE)| ≤ C e−√

π2N/2 (Stenger [2])

|Error (DE)| ≤ C e−(π2N/2)/ log(4πN) (Tanaka et al. [4])

Convergence rates are given (C is not estimated explicitly)

SCAN’2012 8 / 28


Contribution 1. Explicit estimates of the C’s

Case 1

∫ ∞

−∞



There exists a constant C independent of N such that

|Error (SE)| ≤ C e−√

π2N/2 (Stenger [2])

where C = 122.6 (New!)

|Error (DE)| ≤ C e−(π2N/2)/ log(4πN) (Tanaka et al. [4])

where C = 2345 (New!)

Now the errors can be estimated by computing the R.H.S.!

SCAN’2012 9 / 28


Summary table for the explicit estimates

SE/DE-Sinc quadrature

Case 1 (−∞,∞) New!

Case 2 (0,∞) New!

Case 3 (0,∞) New!

Case 4 (a, b) OK [1]

SE/DE-Sinc indefinite integration

Case 1 (−∞,∞) New!

Case 2 (0,∞) New!

Case 3 (0,∞) New!


• The desired explicit estimates are already done in Case 4 [1]

• This study extends the result to Case 1–3

SCAN’2012 10 / 28


Issue on the DE transformation in Case 3

• Inverse of t = ψDE3(x) = ex−exp(−x) cannot be written

explicitly (in elementary functions)

• For the Sinc indefinite integration, its inverse is needed

(not needed for the Sinc quadrature)

• For the purpose, Muhammad–Mori (2003) proposed

t = ψDE3†(x) = log(1 + e(π/2) sinhx)

SCAN’2012 11 / 28


Contribution 2: improving the transformation

• Inverse of t = ψDE3(x) = ex−exp(−x) cannot be written

explicitly (in elementary functions)

• For the Sinc indefinite integration, its inverse is needed

(not needed for the Sinc quadrature)

• For the purpose, Muhammad–Mori (2003) proposed

t = ψDE3†(x) = log(1 + e(π/2) sinhx)

• In this talk, the transformation is improved as

t = ψDE3‡(x) = log(1 + eπ sinhx) (and this is optimal)

SCAN’2012 12 / 28

Outline of this talk

1 Sinc quadrature and Sinc indefinite integration 1

2 Improving the DE transformation in Case 3 13

3 Error estimates with explicit constants 20

4 Summary and future work 26

2 Improving the DE transformation in

Case 3

2 Improving the DE transformation in Case 3

Definition: the strip domain Dd

The integrand should be analytic on the complex domain Dd:

dRe

Im

Dd Dd = {z ∈ C : | Im z| < d}d > 0

Example f(x) =1

1 + x2is analytic on D1

1

Re

Imi

-i

SCAN’2012 14 / 28


Sketch of the existing error analysis [4]

Point discretization error ED and truncation error ET

Let f(ψDE3†(·)) be analytic and bounded on Ddd

Re

Im

Dd

Integrand FDE3†(x) = e−ψDE3†(x)f(ψDE3†(x))ψ′DE3†(x)∫ ∞

0

e−tf(t) dt ≈ h∞∑

k=−∞

FDE3†(kh) ED = O( e−2πd/h)

≈ h

N∑k=−N

FDE3†(kh) ET = O(e−

π4exp(Nh)

)2πd

h≈ π

4eNh ⇒ h =

log(8dN)

NEDE = O

(exp

{−2πdN

log(8dN)

})SCAN’2012 15 / 28


Improvement of the DE transformation



Re

Im

Dd

ψDE3†(t) = log(1 + e(π/2) sinh(t)) → ψDE3‡(t) = log(1 + eπ sinh(t))∫ ∞

0


k=−∞

FDE3‡(kh) ED = O( e−2πd/h)

≈ hN∑

k=−N

FDE3‡(kh) ET = O(e−

π2exp(Nh)

)2πd

h≈ π

2eNh ⇒ h =

log(4dN)

NEDE = O

(exp

{−2πdN

log(4dN)

})SCAN’2012 16 / 28


Is further improvement possible?



Re

Im

Dd

ψDE3‡(t) = log(1 + eπ sinh(t)) → ψDE3∗(t) = log(1 + e2π sinh(t))∫ ∞

0


k=−∞

FDE3∗(kh) ED = O( e−2πd/h)

≈ hN∑

k=−N

FDE3∗(kh) ET = O(e−

π1exp(Nh)

)2πd

h≈ π

1eNh ⇒ h =

log(2dN)

NEDE = O

(exp

{−2πdN

log(2dN)

})SCAN’2012 17 / 28


Not possible (Sugihara 1998)



Re

Im

Dd

ψDE3‡(t) = log(1 + eπ sinh(t)) → ψDE3∗(t) = log(1 + e2π sinh(t))∫ ∞

0


k=−∞

FDE3∗(kh) ED = O( e−2πd/h)/////////////////////

≈ hN∑

k=−N

FDE3∗(kh) ET = O(e−

π1exp(Nh)

)2πd

h≈ π

1eNh ⇒ h =

log(2dN)

NEDE = O

(exp

{−2πdN

log(2dN)

})SCAN’2012 18 / 28


Numerical Example (C, quadruple precision)∫ ∞

0

e−t√1 + tanh2(log(sinh(t))/2) dt = 4arcsinh(1)−

√2(1 + log 2)

1e-35

1e-30

1e-25

1e-20

1e-15

1e-10

1e-05

1

0 10 20 30 40 50 60 70 80 90 100

Err

or

N

SE3

DE3’’

DE3 DE3’

• Almost no difference between ψDE3(t), ψDE3†(t), ψDE3‡(t)?

• For ψDE3‡(t), we have explicit error estimates (next)

SCAN’2012 19 / 28

3 Error estimates with explicit

constants

3 Error estimates with explicit constants

ToDo: Bound the transformed integrand

Case 1

∫ ∞

−∞

f(t)

1 + t2dt =

∫ ∞

−∞

f(ψ(x))

1 + {ψ(x)}2ψ′(x) dx

(ψ(x) = ψSE1(x) or ψDE1(x))

• |f(ψ(z))| ≤ K: Suppose given by users (depends on the problem).

•∣∣∣∣ ψ′(z)

1 + {ψ(z)}2

∣∣∣∣ ≤ ??: Evaluated by this study (always the same)

Existing bound∣∣∣∣ ψ′SE1(z)

1 + {ψSE1(z)}2

∣∣∣∣ ≤ C e−|Re z|,

∣∣∣∣ ψ′DE1(z)

1 + {ψDE1(z)}2

∣∣∣∣ ≤ C e− exp(|Re z|)

The explicit form of C’s is revealed in this study (for Case 1–3), which

enables us to obtain the desired explicit error estimates.SCAN’2012 21 / 28


Error estimates with explicit constants (SE)

Theorem

• f(ψSEi(·)) is analytic on Dd. (i = 1, 2, 3)

• ∀z ∈ Dd, |f(ψSEi(z)| ≤ K. (i = 1, 2, 3)

=⇒|ESE-Sinc quadrature| ≤ KCSECi e

−√2πdN ,

where

CSE = 1 +2

(1− e−√2πd) cos d

, C1 = 22, C2 = 2, C3 = 21/2.

SCAN’2012 22 / 28


Error estimates with explicit constants (DE)

Theorem

• f(ψDEi(·)) is analytic on Dd. (i = 1, 2, 3)

• ∀z ∈ Dd, |f(ψDEi(z)| ≤ K. (i = 1, 2, 3)

=⇒|EDE-Sinc quadrature| ≤ KCDECi e

−2πdN/ log(8dN), (i = 1, 2)

|EDE-Sinc quadrature| ≤ KC ′DECi e

−2πdN/ log(4dN), (i = 3)

where

CDE, C′DE = (only depend on d), C1 = 22, C2 = 2, C5 = 21/2.

SCAN’2012 23 / 28


Numerical Example (C, quadruple precision)

Case 1

∫ ∞

−∞



1e-35

1e-30

1e-25

1e-20

1e-15

1e-10

1e-05

1

0 10 20 30 40 50 60 70 80 90 100

Err

or

N

DE1

error estimate

SE1

error estimate

• “error estimate” bounds the actual error

• Similar results can be obtained in Cases 2 and 3SCAN’2012 24 / 28


Numerical Example (C, double precision)

Case 1

∫ x

−∞

1

1 + t2dt =

π

2+ arctan(x)

1e-16

1e-14

1e-12

1e-10

1e-08

1e-06

1e-04

0.01

1

0 10 20 30 40 50 60 70 80 90 100

Err

or

N

error estimate

error estimate

SE1DE1

• “error estimate” bounds the actual error

• Similar results can be obtained in Cases 2 and 3

SCAN’2012 25 / 28

4 Summary and future work


Error estimates with explicit constants


Case 1 (−∞,∞) New!

Case 2 (0,∞) New!

Case 3 (0,∞) New!



Case 1 (−∞,∞) New!

Case 2 (0,∞) New!

Case 3 (0,∞) New!


t = ψDE3(x) = ex−exp(−x) → t = ψDE3†(x) = log(1 + e(π/2) sinhx)

→ t = ψDE3‡(x) = log(1 + eπ sinhx) (New!)

SCAN’2012 27 / 28


Future work


Case 1 (−∞,∞) New!

Case 2 (0,∞) New!

Case 3 (0,∞) New!



Case 1 (−∞,∞) New!

Case 2 (0,∞) New!

Case 3 (0,∞) New!


• Explicit error estimates for the Sinc approximation (almost done)

• Application to the integral equations over the infinite interval

SCAN’2012 28 / 28

Documents

Error estimates with explicit constants for Sinc quadrature and …conf.nsc.ru/files/conferences/scan2012/139602/Okayama-scan2012.pdf · 1 Sinc quadrature and Sinc inde nite integration